Why AI agents invent their own language if you let them chat

Tom Howarth

doi:10.1126/science.aem6544

AI & machine learningAgents & reasoning

Why AI agents invent their own language if you let them chat

Tom Howarth

Peer-reviewed journalBroad interestEasy to read

In the authors' words

This journal does not let us republish the full abstract. Here are the two sentences Pipette selected, quoted from it. Read the rest at the publisher.

Main result

"Weird terms" are a natural-but disturbing-byproduct of agent swarms.

Read the paper

Appeared: Saturday, September 26. Science. Peer-reviewed journal.

DOI: 10.1126/science.aem6544

AI & machine learning Efficiency & hardware Sep 26, 2026
Double descent without digital computation
Here we demonstrate double descent in a decentralized analog network of self-adjusting resistive elements.
Sam Dillavou, Jason W. Rocks, Jacob F. Wycoff and 2 more — Proceedings of the National Academy of Sciences
Peer-reviewed journalClaims a big step
AI & machine learning Efficiency & hardware Sep 26, 2026
Microrings as programmable temporal kernels enabling photonic AI beyond 100 Gbaud
Here, we redefine MRRs as programmable temporal convolution kernels by exploiting their impulse responses, enabling computation beyond the resonance linewidth.
Shaojie Liu, Tengji Xu, Benshan Wang and 5 more — Nature Communications
Peer-reviewed journalBold claims, read criticallyClaims a big stepReal-world use
AI & machine learning AI for science & medicine Sep 26, 2026
UpTCR: a unified progressive knowledge transfer foundation model for robust T-cell receptor-antigen binding recognition
Here, we present UpTCR, a unified progressive knowledge-transfer foundation model that learns from incomplete data to predict TCR-antigen-HLA binding.
Tianxu Lv, Yang Xiao, Li Chen and 22 more — Nature Communications
Peer-reviewed journalReal-world use
AI & machine learning AI for science & medicine Sep 26, 2026
DR-GEM enables self-supervised machine learning for single-cell embeddings and annotations
Addressing this, we present “Distributionally Robust and latent Group-AwarE consensus Machine learning” or DR-GEM, a self-supervised meta-algorithm that uses the reconstruction error to reorient the model attention to patterns it initially missed and applies balanced consensus learning to increase robustness and filter low-quality data.
Christine Y. Yeh, Min Sun, Livnat Jerby‐Arnon and 1 more — Nature Communications
Peer-reviewed journal

All new papers on Agents & reasoning

Why AI agents invent their own language if you let them chat

In the authors' words

More on Agents & reasoning that day

Double descent without digital computation

Microrings as programmable temporal kernels enabling photonic AI beyond 100 Gbaud

UpTCR: a unified progressive knowledge transfer foundation model for robust T-cell receptor-antigen binding recognition

DR-GEM enables self-supervised machine learning for single-cell embeddings and annotations